Professor of Aeronautics and Astronautics
Director Model-based Embedded and Robotic Systems Group
Computer Science and Artificial Intelligence Laboratory, MIT
MIT Department of Aeronautics and Astronautics
Industry is driving towards rapid manufacturing of systems with increased complexity, while reducing lot size, and re-tooling time and cost. To support these trends, the Model-based Embedded and Robotic Systems (MERS) group at CSAIL, MIT is developing task-executives that enable manufacturing lines to be highly reconfigurable, and enable humans and robots to work together as teams.
The MERS group has a long track record of developing a wide range of systems that are easy for humans to interact with, that are robust and that are highly reconfigurable. We accomplish this through the development of "task executives" for autonomous systems and robots that 1) are commanded in plain English in terms of high-level goals, 2) are able to plan novel ways to achieve these goals, 3) respond intelligently to disturbances encountered when executing these novel plans, and 4) are sensitive to the risk that these novel plans may fail. Applications over the last 25 years range from NASA's Deep Space One probe to autonomous air taxis and manufacturing robots. Additional applications include automobiles, copiers, autonomous air vehicles and submarines, naval ships and smart buildings.
In the context of manufacturing, our goal is to improve the robustness and reconfigurability of manufacturing robots individually and the manufacturing line as a whole. In this talk we demonstrate progress towards three objectives: our first objective is to enable manufacturing robots to perform coordinated tasks robustly in semi-structured environments, and along side humans. In support of this, we use machine-learning technologies to allow robots to learn aspects of new tasks and new configurations on their own, and to learn to perform their tasks more efficiently. Our second objective is to make it simple, fast and cost effective to bring new tasks on line, by reconfiguring robots and the manufacturing line. This is enabled through domain independent, “generative” planning methods that construct novel solutions using composable models. Our final objective is to enable the manufacturing line, as well as individual robots, to optimize and adapt dynamically, for example through rescheduling and reallocation of tasks, while guaranteeing that the risk of task failure is within an operator specified limit. This is enabled through online stochastic methods for “chance-constrained” decision-making. We demonstrate our progress towards each objective in the context of aerospace manufacturing.